تحلیل و پیش‌بینی ترافیک با استفاده از تصاویر مستخرج از Google Maps و شبکه عصبی مکانی - زمانی

The accurate and timely prediction of urban traffic states constitutes a cornerstone of effective route optimization, travel time minimization, and the advancement of intelligent transportation systems (ITS). In megacities, where traffic dynamics exhibit heightened complexity due to intricate spatial interactions and rapidly evolving temporal patterns, there is an imperative for models that jointly capture spatiotemporal dependencies with high fidelity. Conventional fixed sensing infrastructures—such as inductive loop detectors and surveillance cameras—while widely deployed, suffer from substantial installation and maintenance costs, restricted spatial coverage, and reliance on static infrastructure, thereby limiting their scalability and practicality in large-scale urban environments.

This study introduces a novel data-driven spatiotemporal framework tailored for short-term traffic state estimation and forecasting across the extensive highway network of Tehran. The approach harnesses publicly accessible Google Maps traffic imagery as a scalable, low-cost alternative data source, yielding a comprehensive dataset comprising four-level congestion classifications sampled at 15-minute intervals over a three-month period. This large-scale, real-world corpus effectively encapsulates traffic dynamics across diverse congestion regimes throughout the city's highway system. Initial descriptive statistical analyses elucidated pronounced recurrent peak-hour patterns and statistically significant weekday–weekend disparities, establishing a robust empirical foundation for subsequent predictive modeling.

In the core modeling stage, we propose a hybrid spatiotemporal architecture that integrates Graph Convolutional Networks (GCN) with Temporal Convolutional Networks (TCN). The highway network is modeled as a graph, with nodes representing individual highway segments and edges reflecting spatial adjacency. Graph convolution operations effectively capture localized spatial dependencies and congestion propagation effects among neighboring segments, while one-dimensional temporal convolutions model non-stationary temporal evolutions and dynamic traffic trends. In contrast to recurrent architectures, the GCN–TCN paradigm supports fully parallelizable computations, yielding superior training efficiency and markedly reduced computational overhead while preserving intricate spatiotemporal correlations.

The proposed model's predictive efficacy was rigorously assessed across 15-, 30-, and 60-minute horizons using Accuracy and macro F1-score as primary evaluation metrics. Comparative benchmarking against three competitive baselines—Historical Average, standalone Long Short-Term Memory (LSTM), and TCN without explicit spatial modeling—demonstrated consistent superiority of the GCN–TCN framework across all horizons. For the 15-minute forecast, the model attained an Accuracy of 0.70 and F1-score of 0.66, surpassing the strongest baseline (TCN) at 0.65 and 0.64, respectively. At the 30-minute horizon, performance reached 0.68 Accuracy and 0.63 F1-score, outperforming LSTM (0.63/0.58) and TCN (0.64/0.62). Even at the 60-minute horizon, the model sustained strong results with 0.64 Accuracy and 0.60 F1-score. Relative to the Historical Average benchmark, the proposed framework delivered improvements of up to 15% in Accuracy and 16% in F1-score, with particularly pronounced gains under moderate-to-heavy congestion regimes—underscoring enhanced robustness against class imbalance and sudden traffic perturbations.

Collectively, these results affirm that the synergistic integration of cost-effective, image-derived traffic data with graph-informed spatiotemporal deep learning architectures offers a highly efficient, scalable, and economically viable pathway for short-term urban traffic forecasting. The explicit modeling of road network topology emerges as a pivotal factor in elevating predictive precision, laying a solid groundwork for the deployment of real-time traffic systems and advanced decision-support platforms in metropolitan traffic management.

Collectively, these results affirm that the synergistic integration of cost-effective, image-derived traffic data with graph-informed spatiotemporal deep learning architectures offers a highly efficient, scalable, and economically viable pathway for short-term urban traffic forecasting. The explicit modeling of road network topology emerges as a pivotal factor in elevating predictive precision, laying a solid groundwork for the deployment of real-time traffic systems and advanced decision-support platforms in metropolitan traffic management.

Collectively, these results affirm that the synergistic integration of cost-effective, image-derived traffic data with graph-informed spatiotemporal deep learning architectures offers a highly efficient, scalable, and economically viable pathway for short-term urban traffic forecasting. The explicit modeling of road network topology emerges as a pivotal factor in elevating predictive precision, laying a solid groundwork for the deployment of real-time traffic systems and advanced decision-support platforms in metropolitan traffic management.